Data Modeling Of Centrifugal Compressor Based On Rbf Neural Network

Centrifugal compressor which is key equipment in compressing and transporting gases has been widely used in metallurgy, petroleum, chemical and other industrial fields. It plays an important role in industrial production. It has a lot of advantages as well as some insurmountable shortcomings such as narrow steady operating area, surge easily. Surge is the inherent characteristics of the centrifugal compressor, it is harmful to compressor, even endangers the operation safety. Therefore, it is necessary to study the properties of the compressor so as to guarantee the security and highly efficient of the operation. Establishing a precise compressor model is an important way, which has important significance for the study of properties.The centrifugal compressor in gas system of combined cycle power plant of Baosteel was taken as research object in this thesis. In view of the complex mechanism, nonlinear, multivariable and the difficulty in establishing a precise model based on mechanism analysis method, this thesis adopt data-driven modeling method to build up centrifugal compressor model based on radical basis function (RBF) neural network; Whether the application of RBF neural network is successful or not depend on the choice of its model structure largely, and the choice of nodes of hidden layer problem is crucial to approximation performance of RBF neural network. In order to solve the of nodes selection, a thorough analysis of partial least squares(PLS) and fuzzy c-means clustering(FCM) which are methods for solving nodes of RBF neural network was carried out. The co-evolutionary ideology was introduced to combine with FCM, which formed cooperative clustering method. On this basis, a novel two-stage method that consists of cooperative clustering and PLS was proposed. The two-stage method trains a initial RBF neural network with excess nodes by cooperative clustering method, then prunes away the redundant nodes at one time by PLS under the analysis of output information of its hidden nodes and the corresponding optimized model can be obtained in the end.Key variables that affect centrifugal compressor were selected under the analysis of technological process and working mechanism of compressor, and process data was used to establish the model of centrifugal compressor at last. Simulation results show that the two-stage method that consists of cooperative clustering and PLS is a more effective algorithm for solving hidden nodes of RBF neural network, and the constructed model is precise and reliable, implying that the constructed data-driven model can act as an important role in the research of centrifugal compressors.